Back to products

Hakkan

A research and writing assistant that turns what people are saying about a topic into a cited report, then helps you write from it in your own voice.

Own product · Solo: product, design and copy, built with coding agents · Jul 2026 – present

  • Product build
  • AI engineering
  • UX
  • Content discovery

Hakkan (発刊, “to publish”) collects online discussion about a topic into a cited report, then helps you write from it in your own voice.

A Hakkan report: 965 items cited and 903 real voices across seven platforms, with the sentiment split beneath.

Why it exists

Since late 2024, more new articles have been written by AI than by people (Graphite, 2025), and much of it has no real sources. Hakkan starts from real discussions, and the AI model is never used as a source.

How it works

Enter a topic and it builds a visual report of themes and sentiment, with every quote cited. You then write from the report in a voice learned from your own writing.

Where it listens

Searching every source for every question is slow and expensive and returns mostly noise, so Hakkan picks sources for each question.

Per question
Source selectionEach question is split into sub-queries, and only relevant sources are kept.
Then deeper
Specialist sourcesIt adds sources the main search misses, such as review sites for travel and Hacker News for developers.
Yours
Adjustable depthThe user sets how wide and how far back each search goes.

Improving the evidence filter

The first evidence filter removed too many real comments: long articles scored high, while a short reply like “Why need a nanny if I won’t have a job” was marked off-topic. The fix was to judge each comment against the post it replies to.

Human sources in citations

The target was for 40% of citations to come from real people rather than publications. It took four attempts.

StageVoice ratioWhat changed
First measure12%Invalidated: a test flag was replaying cached data instead of searching live.
Live streaming29%Live runs. Better, but the filter still removed real comments.
Facts added19%More article sources lowered the share of human comments. The filter was the bottleneck.
Filter rewritten57%Comments judged in context. Well above the target.

Measured on live runs against a 40% target set before the work started.

Quality rules

Quality rules are enforced in code. Three are built into the product:

3-way
Labelled numbersEvery number is marked as sourced, from your writing, or model-derived. Model-derived numbers are flagged.
0
No virality predictionsWithout outcome data it won't predict virality. It shows what has already performed well instead.
Scoped
Sample sizes statedClaims say “60% of the voices in this report”, never “60% of people”.

A fixed cost limit per run is also enforced in code to keep the service affordable to run.

The screens

  • Hakkan's search screen: one question box, depth and time-range controls, and the platforms it listens across.
  • A finished report: 965 items cited, 903 real voices, themes and sentiment with the receipts one tap away.
  • The visual report, charting where the evidence and the sentiment diverge.
  • Personas: Hakkan trained on your own writing, so the output sounds like you rather than like a model.

Built with

  • Next.js 16
  • React 19
  • TypeScript
  • Tailwind v4
  • Supabase
  • Vercel AI SDK
  • Claude + Grok
  • Paddle
  • Recharts
  • Sentry
  • Claude Code